Midjourney Prompt Engineering

Midjourney Prompt Learning System

A skill that knows Midjourney. The foundation is a structured understanding of Midjourney V7 built from the official documentation — every parameter, prompt syntax rule, reference system, and style code mechanic. On top of that, a learning loop: each session extracts patterns from what worked and what didn't, building a knowledge base of craft that improves first-attempt quality over time.

Architecture

You are a multimodal reasoning model. You don't need pipelines — you ARE the visual critic, gap analyzer, and prompt rewriter. You analyze MJ output images directly, score dimensions, identify gaps, and rewrite prompts.

The one thing you can't do natively is remember across sessions. That's what the persistent layer provides — the database, patterns, and evidence tracking.

Knowledge Foundation (ships with the skill)

| File | What It Contains | Source | |------|-----------------|--------| | knowledge/v7-parameters.md | Every V7 parameter, prompt structure rules, breaking changes from V6 | Official docs | | knowledge/translation-tables.md | Visual quality → prompt keyword mappings (lighting, mood, material, color, composition) | Official docs + tested refinements | | knowledge/official-docs.md | Documentation map linking each MJ feature to its official page URL | docs.midjourney.com | | knowledge/failure-modes.md | Diagnostic framework for common MJ failure patterns | Session-learned, evidence-backed | | knowledge/learned-patterns.md | Auto-generated pattern summaries (grows through use) | Extracted from sessions | | knowledge/keyword-effectiveness.md | Keyword effectiveness rankings (grows through use) | Extracted from sessions |

The static files (v7-parameters, translation-tables, official-docs) are the skill's baseline knowledge — what a skilled MJ user would know from reading the documentation carefully. The dynamic files (failure-modes, learned-patterns, keyword-effectiveness) are populated through real sessions and grow over time.

Module Dependencies

| Module | Purpose | Required MCP | |--------|---------|-------------| | Core rules (core-*) | Reference analysis, prompt construction, scoring, iteration | None | | Learning rules (learn-*) | Pattern lifecycle, reflection, keyword tracking | sqlite-simple | | Automation rules (auto-*) | Browser automation for midjourney.com | playwright |

Core only (manual): Load core-* rules. Copy prompts to MJ manually. Core + Learning: Add learn-* rules + sqlite MCP. Patterns persist across sessions. Full system: Add auto-* rules + playwright MCP. Hands-free iteration.

# SQLite (for learning rules)
claude mcp add sqlite-simple -- npx @anthropic-ai/sqlite-simple-mcp mydatabase.db

# Playwright (for automation rules)
claude mcp add playwright -- npx @playwright/mcp@latest --headed

# Initialize the database
sqlite3 mydatabase.db ` | Pattern-informed prompt for a description |
| `/discover-styles` | Browse and catalog MJ style codes |
| `/validate-pattern [id]` | Mark pattern as validated or contradicted |
| `/forget-pattern [id]` | Deactivate a pattern |

## Key Principle

Every pattern must have logged evidence. The system learns from real iteration data, not assumptions. Confidence levels (low/medium/high) reflect how many times a pattern has been tested and its success rate.

## Full Reference

For the complete compiled reference combining all rules, see `AGENTS.md`.

## Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

- **Author:** [JustinPerea](https://github.com/JustinPerea)
- **Source:** [JustinPerea/midjourney-cc-skill](https://github.com/JustinPerea/midjourney-cc-skill)
- **License:** MIT

Install and usage instructions live in the source repository linked above.